@deepseek-ai/dsh-agent-governance

@deepseek-ai/dsh-agent-governance

English | 中文

One optional Harness Plugin that distributes governance Skills and coordinates Codex, Claude Code, and a child DeepSeek Harness runtime through ctx.subagents.

The plugin mounts the existing filesystem skill provider with an embedded bundled root. Skill catalog summaries remain available through the normal dsh-tool-skill flow; full instructions and referenced resources load only when a skill is selected. Project and user skill roots keep their normal precedence and scope behavior.

The Runtime provides a Provider-backed registry, deterministic recommendations, risk and permission metadata, and model-facing governance_* tools for availability checks, routing, approval, delegation, cancellation, reporting, handoff, and acceptance. Routing never grants execution authority: delegation requires a recorded approval, and a child report is not acceptance.

Governance facts are appended as replayable governance/* Session Events, including Provider checks, approvals, delegation lifecycle, evidence, handoffs, and acceptance. Large task packets remain in workspace files and are referenced by the session rather than copied into every prompt. Handoff paths are restricted to the session workspace. The Plugin does not modify agent-loop; Codex and Claude Code use their existing Providers, while DeepSeek Harness uses the existing subagent-dsh-sdk process boundary when that Provider is loaded.

Delegation uses the tool cancellation signal and a configurable task timeout (30 minutes by default); both cancel the existing Subagent Run and are recorded separately from the child report. The package also exports type-only contributor interfaces for a future native extension bridge. They describe lifecycle seams without registering hooks or changing the Agent Loop in this release.

Mount

Add the package to a profile or bundle patch:

- id: agent-governance
  name: '@deepseek-ai/dsh-agent-governance'

Model Experience

Bundled governance and execution skills

What the model sees

Six short catalog entries from dsh-tool-skill, including runtime, routing, handoff, and acceptance guidance. A selected skill body and its referenced resource guidance appear only after the model loads that skill through the existing skill tool.

Token effect

The initial catalog adds six descriptions; each loaded body adds its own retained tool-result content. The plugin does not duplicate a body into a second synthetic prompt message.

KV Cache effect

The catalog and loaded bodies append model-visible history after the existing reusable prefix. Enabling or disabling this plugin changes the catalog suffix but does not alter unrelated provider behavior.

Known Limitations and Deferred Work

• The Plugin does not modify agent-loop; provider availability still depends on the selected Profile Bundle.
• The bundled resources are versioned with this package; body changes require a package update.
• The filesystem provider may also expose project and user skills when includeDefaultRoots remains enabled.
• Optional helper LLM (helperProvider / helperModel, auto local-edge or Settings): helperRouteHints appends advisory reasons via routeForAsync / routeWithHelperHint and never bypasses approve(); helperContextAssist injects a logged brief through agent.inject (purpose: 'context-assist'). Missing helper follows helperFallback (skip or fail).